Traffic signal adaptive control system and method

Abstract

The invention discloses a traffic signal adaptive control system and method, and the system comprises a sensing and transmission module which collects the multi-mode traffic flow data of an intersection in real time, and compresses and transmits the data; the space-time diagram module abstracts intersections into dynamic space-time diagrams according to the traffic flow data: nodes represent physical lanes, edges are connected with adjacent intersections or upstream and downstream road sections, and weights of the edges are dynamically calculated and updated in real time according to historical steering probabilities, phase constraint influence factors and real-time queuing length influence factors; the decision module is used for generating an action space containing a phase switching time sequence according to the state space on the basis of the space-time diagram, so that the number of vehicles passing in unit time is maximum, and the maximum waiting time difference in each direction is minimum; and the phase dynamic optimization module receives the action space, responds to the conventional optimization trigger condition, dynamically adjusts the phase green light duration and the switching sequence, and can solve the problems of long-term waiting of vehicles on the next trunk road and reduced traffic efficiency at the intersection in the tidal flow or emergency scene.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic control systems, and in particular to a traffic signal adaptive control system and method. Background Art

[0002] In urban smart transportation management systems, the dynamic optimization of traffic signal configuration has become a core component for improving road network efficiency. Existing adaptive signal control systems primarily rely on a single data source or static optimization, lacking real-time data-driven capabilities and struggling to cope with the spatiotemporal dynamics of traffic flow.

[0003] For example, there is a Chinese patent with publication number CN104040605B, which involves a traffic signal control method and a traffic signal controller, which realizes the control of traffic signals. However, this patent ignores the topological relationship between upstream and downstream road networks, such as turning probability and queue overflow effect, and is prone to falling into the local optimal trap. It only optimizes single-point congestion, aggravating secondary queues at related intersections, and its phase decision does not integrate historical traffic patterns and cannot adapt to scenarios such as tidal flow; it takes "maximizing network throughput" as the single goal and ignores the balance of lane-level waiting time, which will lead to a significant increase in vehicle delays on secondary roads when traffic is unbalanced. Summary of the Invention

[0004] In order to solve the problems of long waiting times for vehicles on secondary roads and reduced intersection efficiency caused by the lack of vehicle dynamic intention perception and insufficient upstream and downstream spatiotemporal correlation in tidal flow or emergency scenarios, the present invention proposes a traffic signal adaptive control system and method. By quantifying the coupling effect of historical turning probabilities and real-time phase conflicts through a spatiotemporal graph, and combining it with dual-objective optimization, it achieves adaptive adaptation to changes in traffic demand.

[0005] A further object of the present invention is to prevent green time from being wasted or insufficient by dynamic phase splitting and priority rearrangement.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a traffic signal adaptive control system, comprising: The perception and transmission module collects multimodal traffic flow data at intersections in real time and compresses and transmits it; The spatiotemporal graph module abstracts intersections into dynamic spatiotemporal graphs based on traffic flow data. Nodes represent physical lanes, and edges connect adjacent intersections or upstream and downstream sections. Edge weights are dynamically calculated and updated in real time based on historical turn probabilities, phase constraint influencing factors, and real-time queue length influencing factors. The decision module, based on the space-time graph, generates an action space containing the phase switching sequence according to the state space, so as to maximize the number of vehicles passing per unit time and minimize the difference in the maximum waiting time in each direction; The phase dynamic optimization module receives the action space, responds to the conventional optimization trigger conditions, and dynamically adjusts the phase green light duration and switching sequence.

[0007] In this technical solution, full-domain real-time data collection and ultra-low latency, high-compression transmission are achieved through the perception and transmission modules; dynamic topology modeling is achieved through the space-time graph module. The space-time graph module introduces a dynamic space-time graph for the first time to characterize intersections. The weights of the edges are dynamically adjusted, and the impact of phase conflicts and real-time queue lengths is taken into account to improve traffic efficiency and control flexibility; the decision-making module adopts dual-objective optimization to simultaneously meet the maximum number of vehicles passing per unit time and the minimum difference in the maximum waiting time in each direction, taking into account both efficiency and waiting time balance; millisecond-level green light adjustment is achieved through the phase dynamic optimization module.

[0008] Preferably, the phase dynamic optimization module includes: responding to emergency event requests, inserting priority phases and locking the red light in the conflict direction; when a V2X emergency event is detected, interrupting the current service program, locking the green light in the priority direction and the red light in the conflict direction within the response time; the response time is less than 1 second, introducing an emergency response mechanism without manual processing, and improving emergency passage efficiency and safety; executing at least one dynamic adjustment strategy of phase splitting or merging or skipping or priority reordering; making the system "immune to mutations" and quickly completing signal reconfiguration in emergency scenarios such as accidents and rescue, avoiding the spread of congestion, and shortening emergency response time.

[0009] Preferably, the weight of the edge is obtained by weighting the historical turning probability, the phase constraint influence factor and the real-time queue length influence factor with their respective weight coefficients and then summing them up: the weight coefficient of the historical turning probability is dynamically adjusted according to the emergency; the weight coefficient of the phase constraint influence factor is dynamically adjusted according to the phase conflict matrix; the weight coefficient of the real-time queue length influence factor is dynamically adjusted according to the current traffic; the weight adaptive mechanism allows the space-time graph to quickly complete topology reconstruction in sudden change scenarios such as accidents and lane closures, thereby improving the real-time and accuracy of the control strategy.

[0010] Preferably, the conventional optimization triggering conditions are: when the lane saturation is greater than a first threshold or the queue length difference is greater than a second threshold, the green light duration is adjusted; daily congestion self-recovery is achieved through a dual-threshold triggering strategy, for example, when the saturation is greater than 80% or the queue length difference is greater than 40%, the bottleneck phase green light is automatically extended to improve peak-hour traffic capacity.

[0011] Preferably, the dynamic adjustment strategy includes: splitting the composite phase into independent phases when the traffic in one direction surges; inserting a same-release phase into the symmetrical release phase; skipping the phase when there is no vehicle passing for multiple consecutive cycles; merging multiple independent phases into a composite phase when there are serious empty phases, maximizing the green light utilization rate through flexible phase control, reducing the empty rate, reducing signal cycle loss, and reducing the average vehicle delay during off-peak hours.

[0012] Preferably, the dynamic adjustment strategy includes: when the flow rate in the tidal flow direction accounts for more than a third threshold, increasing the phase priority in that direction and extending the green light duration; increasing the priority of the high-demand phase to improve the flexibility and rationality of resource flow.

[0013] Preferably, in the spatiotemporal graph module, the attributes of the nodes include current flow, average speed and saturation, and the spatiotemporal graph module reconstructs the graph structure at regular intervals to update the real-time status.

[0014] Preferably, the perception and transmission module uses the Kalman filter algorithm to perform drift correction on the high-precision positioning data collected by the on-board terminal, and matches it with the high-precision map to lane-level accuracy to ensure the real-time and accuracy of vehicle dynamic data.

[0015] Preferably, in the decision module, the state space includes the current phase combination, the number of vehicles in each direction, the saturation in each direction and the remaining time of the phase, and the action space includes the green light duration adjustment amount, the red light duration adjustment amount and the phase priority adjustment coefficient.

[0016] The present invention also adopts the following technical solution: a traffic signal adaptive control method, based on the above-mentioned traffic signal adaptive control system, comprising the following steps: S1, collects vehicle dynamic data, lane-level traffic parameters and historical traffic flow data in real time to build a perception network; S2, constructs a dynamic spatiotemporal graph based on the perception network and the current phase state; S3, generates signal control strategies based on the state data of the spatiotemporal graph; S4, execute the strategy and adjust the traffic light, and feed the execution effect back to the perception network for re-collection.

[0017] The beneficial effects of the present invention are: 1) Build an integrated vehicle-road-cloud architecture to break through the limitations of perception dimensions and achieve accurate prediction of unexpected behaviors; 2) Dynamic spatiotemporal graph modeling is adopted, nodes are dynamically reconstructed, and edge weights are adaptively adjusted based on historical turning probabilities, phase conflict influencing factors, and real-time queuing influencing factors, eliminating local optimization traps and improving traffic efficiency; 3) Adopting dynamic phase adjustment strategy, low-flow phases are automatically skipped and high-flow phases are extended with green light, thus reducing signal cycle loss and improving emergency response speed; 4) Adopting dual-objective optimization of traffic efficiency and wait time balance to maintain lane-level wait time balance and prevent increased delays on secondary roads when traffic imbalance occurs; 5) Achieve high-precision regional-level traffic signal adaptive control while balancing the contradiction between global collaborative optimization and local real-time response. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the method in Example 2 of the present invention.

[0019] Figure 2 This is a phase pattern diagram of the intersection signal solution in Example 2 of the present invention.

[0020] Drawing numbers: north-south entrance go straight 1, north-south entrance turn left 2, east-west entrance go straight 3, east-west entrance turn left 4, north-south entrance go straight and north-south entrance turn left 5, south-south entrance go straight and north-south entrance turn left 6, east-south entrance go straight and north-south entrance turn left 7, west-south entrance go straight and north-south entrance turn left 8. DETAILED DESCRIPTION

[0021] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0022] Example 1 This embodiment provides a traffic signal adaptive control system for achieving millisecond-level traffic light control in complex scenarios such as congestion, imbalance, tidal flow, and emergencies. The control system includes a perception and transmission module, a spatiotemporal graph module, a decision module, and a phase dynamic optimization module. All modules are interconnected via a 5G NR V2X network, achieving end-to-end latency of no more than 10ms and a peak rate of up to 100Mbps.

[0023] The perception and transmission module includes a perception unit and a collection unit.

[0024] The collection unit is composed of three layers: vehicle side, road side and cloud side.

[0025] The vehicle side collects the vehicle's high-precision positioning, instantaneous speed, turn signal status and autonomous driving decision intention in real time through the on-board terminal OBU at a frequency of no less than 10Hz. The data is first filtered by Kalman to eliminate positioning drift, and then matched to the specific lane with the high-precision map.

[0026] RSU radar and video fusion equipment is installed on the roadside on signal poles or gantries to continuously output the number of vehicles passing, arrival rate, queue length and headway of each lane; geomagnetic coils serve as a supplement to provide occupancy data to make up for the shortcomings of video detection in bad weather.

[0027] The cloud accesses the historical traffic database, summarizes long-term traffic flow characteristics at a granularity of 1 minute or 5 minutes, and pulls weather information such as visibility and precipitation intensity in real time; all time series features are mapped to a unified vector space through the Time2Vec encoder to facilitate deep model processing.

[0028] To reduce the channel load, the transmission unit uses a wavelet transform compression algorithm at the sending end to compress the original data volume by at least 80%, and then broadcasts it to the entire network through 5G NR V2X to achieve millisecond-level data sharing.

[0029] The spatiotemporal graph module abstracts the intersection into a dynamic spatiotemporal graph.

[0030] Nodes correspond to physical lanes, with attributes including current traffic volume, average speed, and saturation. Edges are divided into spatial and temporal edges. Spatial edges connect upstream and downstream lanes, describing the physical connectivity of where vehicles originate and where they go. Temporal edges connect lanes in adjacent time steps, depicting the evolution of traffic flow over time.

[0031] The edge weight is no longer simply marked as conflicting or not, but is dynamically calculated based on the historical turn probability, phase constraint influencing factors, and real-time queue length influencing factors, as shown below.

[0032] The weight coefficient of the historical turning probability is 0.7 when there is no emergency, and drops to below 0.3 when an emergency occurs; the phase constraint influence factor is 1 in the green light phase and 0 in the red light phase, and can be temporarily increased to 1.5 when the vehicle has priority; the real-time queue length influence factor is 0.3 in off-peak hours and 0.6 in peak hours.

[0033] The entire graph structure is reconstructed at fixed time intervals (e.g., 15 seconds), and node attributes and edge weights are updated in real time to ensure a response to traffic conditions within seconds.

[0034] The decision module generates a signal control strategy based on the space-time graph.

[0035] The state space includes the current phase combination, the number of vehicles in each direction, the saturation of each direction, and the remaining time of the phase; the action space includes the green light duration adjustment amount, the red light duration adjustment amount, and the phase priority adjustment coefficient.

[0036] The reward function takes into account both traffic efficiency and waiting time balance: traffic efficiency is measured by the ratio of the number of vehicles passing through a single cycle to the cycle length; waiting time balance is measured by the normalized difference between the maximum waiting time and the average waiting time in each direction, and the goal is to minimize this difference.

[0037] Deep reinforcement learning algorithms (such as PPO combined with spatiotemporal Transformer) output optimal actions within milliseconds, enabling real-time optimization of traffic lights.

[0038] The phase dynamic optimization module is responsible for converting strategies into executable signals.

[0039] The conventional trigger mechanism is evaluated after each signal cycle. If the saturation of any lane exceeds 80% or the difference in queue length exceeds 40%, the green light duration is adjusted according to the optimization results.

[0040] Upon receiving an emergency event request (such as accident, rescue, or bus priority) from the onboard V2X, the emergency response mechanism immediately interrupts the current program, inserts a priority phase, and keeps the red light in the conflicting direction locked.

[0041] The dynamic strategy library supports phase splitting, merging, skipping, and priority reordering: when traffic in a certain direction of a composite phase surges, it can be split into independent phases; when no vehicles pass through for multiple consecutive cycles, the phase can be skipped to reduce the number of unused green lights; when multiple independent phases are all unused, they can be merged into a composite phase; high-demand phases (such as tidal flow directions) output by reinforcement learning can be automatically released in advance.

[0042] The entire control process forms a closed loop of "real-time perception - intelligent decision-making - precise control". The execution effect is fed back in real time by the perception and transmission modules for the next cycle graph reconstruction and strategy update.

[0043] The following is a detailed description of the system's operation process.

[0044] The global perception network collects and compresses data in real time and transmits it to the cloud via 5G NR V2X. The spatiotemporal map module reconstructs traffic maps in seconds. The decision module generates strategies in milliseconds. The phase dynamic optimization module performs signal adjustments and feeds the results back to the perception network, enabling adaptive management of traffic signals throughout the entire process.

[0045] The traffic signal adaptive control system of this embodiment has dynamic topology reconstruction capabilities and millisecond-level response speeds. It can effectively cope with complex scenarios such as congestion, imbalance, tidal flow, and emergencies. It provides a scalable decision-making framework for intelligent transportation systems and promotes the paradigm shift of urban road signal control from "fixed timing" to "intelligent regulation."

[0046] Example 2 This embodiment provides a traffic signal adaptive control method, based on the above-mentioned traffic signal adaptive control system, such as Figure 1 As shown, the following steps are included.

[0047] Step S1: Construction of the vehicle-road-cloud collaborative perception network.

[0048] Deploy on-board terminals (OBUs), roadside sensing units (RSUs) and geomagnetic coils for multi-source data collection, collecting vehicle dynamic data, lane-level traffic parameters and historical traffic flow data in real time.

[0049] The collected vehicle dynamic data include position, speed and acceleration, lane-level traffic parameters include occupancy, queue length and headway, and historical traffic flow data include historical data for the same period.

[0050] The system constructs a three-dimensional perception network through three types of data sources, as follows.

[0051] Vehicle-side dynamic data: V2X communication is used to obtain real-time high-precision vehicle positioning (GPS / Beidou), instantaneous speed, turn signal status, and autonomous driving decision-making intentions, such as lane changes and braking. The sampling frequency is ≥10Hz. After the Kalman filter algorithm eliminates positioning drift, it is matched with the high-precision map to the specific lane.

[0052] Roadside perception data: By deploying radar and vision fusion equipment, video streams are used to analyze the number of vehicles passing each lane, arrival rate, queue length, and headway.

[0053] Cloud-based collaborative data: Integrates historical traffic databases, weather information, road network structures, etc., and maps time series features into vector space through the Time2Vec encoder.

[0054] Among them, the historical traffic database is integrated according to a certain time granularity, such as 1 minute, 5 minutes, etc.; weather information includes but is not limited to visibility and precipitation intensity.

[0055] In this embodiment, the data transmission mechanism can adopt 5G NR V2X communication technology to implement data transmission, support a peak rate of 100Mbps, end-to-end delay ≤10ms, and a data compression algorithm based on wavelet transform to compress the original data volume by more than 80%.

[0056] Step S2: abstract the intersection into a dynamic spatiotemporal graph.

[0057] The nodes of the dynamic space-time graph represent "physical lanes", which divide the intersection into specific lanes, such as the left-turn lane at the east entrance, the straight lane at the west exit, and so on.

[0058] Edges include spatial edges and temporal edges. Spatial edges connect upstream and downstream lanes, reflecting the physical connectivity of "where the car comes from and where it goes". Temporal edges connect the same lane at time steps t and t+1, representing the evolution of traffic flow over time.

[0059] In the dynamic space-time graph, the weight of the edge is no longer a simple conflict or non-conflict, but a combination of historical turning probabilities, current signal phases, and dynamic values ​​of real-time traffic flow, speed, and saturation, which can be used to predict or optimize the traffic status at the next moment.

[0060] To characterize the traffic turning relationship and phase conflict rules, the edge weight is determined by the historical turning probability of the vehicle's driving trajectory and the signal control constraint, and the edge weight is dynamically updated in real time according to the current phase state.

[0061] A turn relationship refers to the path a vehicle takes from one entry lane (upstream node) to another exit lane (downstream node). For example, a turn relationship from the left-turn lane at the east entrance to the through lane at the north exit is from "east" to "north." In the intersection's dynamic spatiotemporal graph, a turn relationship is represented as a spatial edge, pointing from the upstream node to the downstream node.

[0062] Phase conflict refers to the conflict of vehicle trajectories that exists at the same time or in adjacent time periods due to different signal phases in traffic signal control. Different signal phases are combinations of green lights in different directions, which are specifically manifested as physical conflicts and signal conflicts.

[0063] Physical conflicts occur when traffic flows from different phases overlap, such as when they intersect or merge, potentially leading to collisions. For example, when an eastbound left turn (green light) and a northbound through-traffic (green light) are both released simultaneously, left-turning vehicles must cross through the through-traffic, creating an intersection conflict. When northbound through-traffic is released simultaneously, vehicles may be stranded at the intersection if a full-red time-out period is not set.

[0064] Signal conflict occurs when traffic flows that haven't cleared the previous phase of a signal change conflict with traffic being released in the next phase. For example, after the east-west through-the-road green light ends, the left-turn phase begins immediately. However, through-the-road vehicles may not have fully cleared the intersection due to queues, causing a conflict with left-turning vehicles.

[0065] Phase conflicts can be characterized by a conflict matrix. The conflict matrix is ​​a binary table that marks which phases cannot be released simultaneously. A value of 1 indicates conflict, and a value of 0 indicates compatibility.

[0066] Historical turn probability refers to the probability of a certain turn relationship occurring based on past statistics. For example, past statistics show that the probability of a turn relationship from the east exit to the north exit is 65%. Historical turn probability can determine the potential conflict probability, and signal control constraints can be reflected in the phase conflict matrix.

[0067] The current phase state refers to the set of green light phases that the traffic signal is currently releasing. A traffic signal adaptive control method of the present invention also dynamically updates the edge weights in real time based on the current phase state. For example, when the current phase is green, the edge weight of the corresponding flow direction is increased, and according to the conflict matrix, the conflicting phase weight is reset to zero, so that the associated edge weights are suppressed, avoiding the algorithm from selecting a signal strategy that may cause a conflict.

[0068] For example, when the light for a left turn from north to south is green, the weight of the corresponding left-turn edge increases, and the weight of the edge in the conflicting phase release flow direction is reset to zero.

[0069] Specifically, in this embodiment, the comprehensive weight of the edge is obtained by weighting the historical turning probability, the phase constraint influence factor, and the real-time queue length influence factor by their respective weight coefficients and summing them up.

[0070] The three weight coefficients generally have values ​​between 0 and 1.

[0071] The weight coefficient of the historical turning probability can be adaptively adjusted according to the scenario. For example, when no event occurs, it relies more on the historical turning pattern, and the weight coefficient of the historical turning probability can be taken as 0.7. In the case of an emergency, the weight coefficient of the historical turning probability is reduced to below 0.3 to weaken the impact of historical data and give priority to responding to real-time changes.

[0072] The weight coefficient of the phase constraint influencing factor is 1 during the green light period of the phase, opening the corresponding turn, and 0 during the red light period of the phase, completely blocking the turn. When the vehicle has priority, the value is 1.5, temporarily increasing the weight to speed up the clearance.

[0073] The weight coefficient of the real-time queue length influencing factor is adaptively adjusted according to the scenario. Specifically, the queue length during off-peak hours has little impact on the intersection, and the weight coefficient of the real-time queue length influencing factor can be 0.3; the queue length during peak hours reflects the congestion situation at the intersection, and the weight coefficient of the real-time queue length influencing factor can be 0.6.

[0074] Step S3: Generate a signal control strategy based on the deep reinforcement learning framework according to the spatiotemporal graph.

[0075] The state space should contain key information in each direction that can affect the efficiency of traffic signal control.

[0076] Taking into account the key features of intersection signal control and the road network structure, this embodiment defines the state space of intelligent control of intersection traffic signals as a set including the current phase combination, the number of vehicles in each direction, the saturation of each direction, and the remaining time of the phase.

[0077] In order to better guide vehicles at intersections, the traffic signal control system needs to select the appropriate phase to guide traffic based on the current intersection environment.

[0078] When designing the action space, an action space for intelligent control of intersection traffic signals consisting of a phase pattern is selected.

[0079] Based on the current state, the agent chooses to continue the current phase or switch to the next phase with higher priority.

[0080] Here, the action space is defined as a set including the green light duration adjustment amount, the red light duration adjustment amount, and the phase priority adjustment coefficient.

[0081] Taking the intersection as an example, excluding the right turn, there are eight phase modes for the intersection signal scheme, namely, north-south entrance straight 1, north-south entrance left turn 2, east-west entrance straight 3, east-west entrance left turn 4, north-south entrance straight and north-south entrance left turn 5, south-south entrance straight and north-south entrance left turn 6, east-south entrance straight and north-south entrance left turn 7, west-south entrance straight and north-south entrance left turn 8, as shown in the following figure. Figure 2 shown.

[0082] Then a dual-objective optimization function is designed to balance traffic efficiency and waiting time balance. Traffic efficiency refers to the number of vehicles passing through per unit time, and waiting time balance refers to the maximum waiting time difference in each direction.

[0083] In this embodiment, two objective functions are adopted: traffic efficiency and waiting time balance, as shown in the following details.

[0084] The traffic efficiency is obtained by the ratio of the cumulative number of vehicles passing through N phases in a single cycle to the total duration of the signal cycle. The number of vehicles during the green light period of each lane is counted in real time through roadside sensing equipment, and the green light duration is dynamically adjusted to reduce the time when the green light is left idle.

[0085] The waiting time balance is obtained by normalizing the difference between the maximum waiting time of each lane and the average waiting time of all lanes by dividing it by the average waiting time of all lanes. The cumulative waiting time of each vehicle is obtained through the vehicle-side V2X, and the waiting time of vehicles in all directions is balanced, thereby balancing the vehicle delays in all directions.

[0086] Step S4: dynamic phase optimization logic.

[0087] Real-time optimization trigger mechanisms include regular optimization and emergency response.

[0088] Conventional optimization means starting strategy evaluation after each signal cycle. When lane saturation > 80% or queue length difference > 40% is detected, the green light duration adjustment of the signal scheme is triggered.

[0089] Emergency response refers to the immediate initiation of an interrupt service routine upon receiving an emergency event (such as an accident, rescue, or other priority request) from the vehicle's V2X, inserting a priority release phase or maintaining the green light on the priority phase and locking the red light on the conflicting direction.

[0090] The dynamic adjustment strategy of phase sequence includes phase splitting, phase merging and skipping, phase time optimization and priority reordering. The dynamic adjustment strategy of phase sequence is described in detail below.

[0091] When traffic surges in one direction during a composite phase (e.g., north-south straight-through + left-turn), the traffic flow can be split into independent phases for separate control. Alternatively, a simultaneous release phase (south exit simultaneous release) can be inserted into symmetrical release phases (e.g., north-south straight-through and north-south left-turn) to increase traffic efficiency. If a phase has no vehicles passing or queuing for several consecutive cycles, skip that phase to reduce green light idleness. Alternatively, if multiple independent phases have few vehicles passing and are severely idle, the independent phases can be combined into a composite phase to reduce cycle loss.

[0092] According to the traffic flow, the green light release time of each phase is adaptively adjusted to reduce the queuing time in the high-traffic phase and the idle time in the low-traffic phase.

[0093] Based on the reinforcement learning output ranking, high-demand phases (such as tidal flow direction) are automatically adjusted.

[0094] Phase dynamic optimization logic realizes full-process adaptive management of traffic signal control through a closed-loop mechanism of real-time perception-intelligent decision-making-precise control.

[0095] The beneficial effects of phase dynamic optimization include the following three aspects.

[0096] First, the dual trigger mechanism: integrating regular traffic threshold warnings with emergency event responses, building a seamless system from daily optimization to emergency response.

[0097] Second, flexible phase control: Through a combination of strategies such as phase splitting and merging, duration optimization, and priority reordering, it dynamically balances traffic efficiency and resource utilization, reducing the error in green light time allocation by 25%-40%.

[0098] Third, multi-objective collaborative optimization: Combining reinforcement learning with real-time data fusion technology, it reduces queue differences (<15%) and reduces the number of empty green lights (empty rate <5%), while increasing peak-hour traffic capacity by more than 30%.

[0099] A traffic signal adaptive control method in this embodiment collects multimodal data in real time through a global perception network, transmits it to a large cloud-based model platform via a communication network, models the spatiotemporal characteristics of traffic flow based on a deep prediction model of spatiotemporal Transformer, and generates a dynamic signal control strategy in combination with a reinforcement learning algorithm, which can cope with the dynamic changes in complex traffic scenarios.

[0100] Example 3 To verify the effectiveness of the technical solution of this invention, a typical urban intersection was selected as an implementation case. Located in the city's core area, this intersection is surrounded by commercial areas, residential areas, and schools. Traffic flow is high and complex, with significant tidal flow during peak hours in the morning and evening, and congestion during off-peak hours due to unexpected events.

[0101] Step S1: Build a vehicle-road-cloud collaborative perception network.

[0102] At this intersection, 100 onboard vehicles (OBUs), five roadside sensing units (RSUs), and eight geomagnetic coils were deployed. During the morning rush hour, the OBUs, sampling at a 10Hz rate, use V2X communication to obtain real-time data such as vehicle positioning, instantaneous speed, and turn signal status.

[0103] For example, when a vehicle approaches an intersection, its positioning data is processed by the Kalman filter algorithm and accurately matched to the left-turn lane of the east entrance, providing accurate vehicle dynamic information for subsequent traffic flow analysis.

[0104] The roadside radar and vision fusion equipment analyzes the video stream in real time and calculates parameters such as the number of vehicles passing through each lane and the arrival rate.

[0105] For example, in a certain 5-minute period, the number of vehicles passing through the through lane of the west entrance is 200, the arrival rate is 40 vehicles per minute, and the queue length reaches 50 meters.

[0106] At the same time, cloud-based collaborative data collected historical traffic data for the same time period over the past week, as well as real-time weather information (visibility was good and there was no precipitation that day). These time series features were mapped into a vector space using the Time2Vec encoder, providing comprehensive data support for subsequent modeling and decision-making.

[0107] By utilizing 5G NR V2X communication technology and combining it with the wavelet transform data compression algorithm, a large amount of collected raw data can be efficiently transmitted to the cloud. The transmission delay is controlled within 10ms, and the data volume is compressed by more than 80%, ensuring the real-time and efficient transmission of the data.

[0108] Step S2: spatiotemporal graph network modeling.

[0109] The intersection is abstracted into a dynamic spatiotemporal graph, with each physical lane as a node.

[0110] Taking the left-turn lane at the east entrance during the morning rush hour as an example, the attributes of this node include information such as current traffic volume (15 vehicles passing through per minute), average speed (15 km / h), and saturation (0.7).

[0111] Edges connect adjacent intersections and upstream and downstream sections of the same road, reflecting the temporal evolution pattern.

[0112] The edge weight is calculated by summing the historical turning probability, phase constraint influence factor, and real-time queue length influence factor weighted by their respective weight coefficients.

[0113] During the normal morning rush hour, if no emergencies occur, the weight coefficient of the historical turning probability is 0.7, and the historical turning probability shows that the probability of turning left from the east to the north exit is 60%; the current north-south left turn is green, and the weight coefficient of the phase constraint influence factor is 1, and the phase constraint influence factor is 1; the real-time queue length influence factor is calculated based on the queue length, and the weight of the real-time queue length influence factor is 0.6.

[0114] The weight of the edge is obtained through comprehensive calculation and used for subsequent signal control decisions.

[0115] Every 15 seconds, the graph structure is reconstructed based on real-time traffic status changes.

[0116] For example, when the queue length of the through lane at the west entrance suddenly increases, the attributes and edge weights of the corresponding nodes and edges will be updated in a timely manner to accurately reflect the changes in traffic status.

[0117] Step S3: multi-objective optimization decision.

[0118] In the traffic signal control at this intersection, the state space includes the current phase combination, the number of vehicles in each direction, the saturation of each direction, and the remaining time of the phase. The current phase combination is a green light for north-south straight driving. The number of vehicles in each direction is counted by roadside sensing equipment. For example, there are 30 vehicles waiting at the east entrance and 25 vehicles waiting at the west entrance. The saturation of each direction is calculated based on the traffic flow and lane capacity, and the remaining time of the phase is updated in real time.

[0119] The action space is a set of green light duration adjustment amounts, red light duration adjustment amounts, and phase priority adjustment coefficients. At a certain moment, based on the current state, the intelligent agent finds that there are many left-turning vehicles at the east entrance and the saturation is high. It decides to increase the green light duration of the left-turn phase of the east entrance by 5 seconds, and correspondingly reduce the red light duration by 5 seconds. At the same time, the phase priority coefficient is adjusted to increase the priority of the left-turn phase of the east entrance.

[0120] The reward function optimizes the signal control strategy by balancing traffic efficiency and wait time. During a signal cycle, roadside sensing equipment counts the number of vehicles passing through each lane during the green light period to calculate traffic efficiency.

[0121] For example, within a certain cycle, 150 vehicles pass through the north-south straight phase and 80 vehicles pass through the east-west left-turn phase. The total signal cycle duration is 120 seconds. The traffic efficiency is (150+80)÷120≈1.92 vehicles / second.

[0122] The vehicle-side V2X system collects the cumulative waiting time of each vehicle and calculates a waiting time balance index. For example, if the average waiting time for all lanes is 30 seconds and the maximum waiting time for a particular lane is 45 seconds, the waiting time balance index is (45 - 30) ÷ 30 = 0.5. By continuously adjusting the action space and optimizing the reward function, intelligent traffic signal control is achieved.

[0123] Step S4: dynamic phase optimization logic.

[0124] In terms of regular optimization, strategy evaluation is performed after each signal period.

[0125] During the evening rush hour, it was detected that the saturation of the east entrance lane reached 85%, triggering an adjustment to the green light duration of the signal scheme.

[0126] Based on real-time traffic conditions, the green light duration for the left-turn phase at the east entrance was increased from 30 seconds to 35 seconds, effectively reducing vehicle queuing time.

[0127] In terms of emergency response, when the on-board V2X sends a request that an ambulance needs urgent passage, the system immediately starts the interrupt service program, inserts the priority release phase, keeps the green light in the direction of the ambulance's travel, locks the red light in the conflicting direction, and ensures that the ambulance passes through the intersection quickly, with the response time controlled within 1 second.

[0128] In the dynamic phase and sequence adjustment strategy, when the left-turn and through-traffic traffic at the north entrance surges during the morning rush hour, the original composite phase (north entrance through + left turn) is split into independent phases for separate control. At the same time, a simultaneous release phase for the south entrance is inserted into the symmetrical release phases (north-south through and north-south left turn), increasing the traffic efficiency in this direction by 20%.

[0129] If no vehicles pass through a phase for several consecutive cycles during off-peak hours, that phase is skipped, reducing the number of empty green lights and signal cycle losses. The green light duration for each phase is adaptively adjusted based on traffic flow.

[0130] For example, during the morning peak tidal flow period, the green light duration for the phase entering the city is increased, while the green light duration for the phase leaving the city is reduced, which effectively reduces the queuing time in the high-flow phase and the idle time in the low-flow phase, reducing the green light time allocation error by about 30%.

[0131] By sorting the outputs based on reinforcement learning, the priority of the high-demand phases in the tidal flow direction is automatically adjusted. For example, when the traffic in the direction of entering the city is heavy during the morning rush hour, the priority of the phase in the direction of entering the city is increased, which increases the traffic capacity during peak hours by 35%, reduces the queue difference to within 10%, and controls the green light vacancy rate within 3%.

Claims

1. A traffic signal adaptive control system, characterized in that: include: The perception and transmission module collects multimodal traffic flow data at intersections in real time and compresses and transmits it; The spatiotemporal graph module abstracts intersections into dynamic spatiotemporal graphs based on traffic flow data. Nodes represent physical lanes, and edges connect adjacent intersections or upstream and downstream sections. Edge weights are dynamically calculated and updated in real time based on historical turn probabilities, phase constraint influencing factors, and real-time queue length influencing factors. The decision module, based on the space-time graph, generates an action space containing the phase switching sequence according to the state space, so as to maximize the number of vehicles passing per unit time and minimize the difference in the maximum waiting time in each direction; The phase dynamic optimization module receives the action space, responds to the conventional optimization trigger conditions, and dynamically adjusts the phase green light duration and switching sequence.

2. A traffic signal adaptive control system according to claim 1, characterized in that: The phase dynamic optimization module includes: responding to emergency event requests, inserting a priority phase and locking a red light in the conflicting direction; and executing at least one dynamic adjustment strategy of phase splitting, merging, skipping, or priority reordering.

3. The traffic signal adaptive control system according to claim 1, characterized in that: The weight of the edge is obtained by summing the historical turning probability, phase constraint influence factor and real-time queue length influence factor weighted by their respective weight coefficients: the weight coefficient of the historical turning probability is dynamically adjusted according to the emergency event; the weight coefficient of the phase constraint influence factor is dynamically adjusted according to the phase conflict matrix; the weight coefficient of the real-time queue length influence factor is dynamically adjusted according to the current traffic.

4. The traffic signal adaptive control system according to claim 1, characterized in that: The conventional optimization triggering condition is: when the lane saturation is greater than a first threshold or the queue length difference is greater than a second threshold, the green light duration is adjusted.

5. The traffic signal adaptive control system according to claim 2, characterized in that: The dynamic adjustment strategy includes: splitting a composite phase into independent phases when traffic in one direction surges; inserting a simultaneous release phase into a symmetrical release phase; skipping a phase when no vehicles pass through for multiple consecutive cycles; and merging multiple independent phases into a composite phase when there is a serious empty release.

6. A traffic signal adaptive control system according to claim 2 or 5, characterized in that: The dynamic adjustment strategy includes: when the flow rate in the tidal flow direction accounts for more than a third threshold, increasing the phase priority in that direction and extending the green light duration.

7. The traffic signal adaptive control system according to claim 1, characterized in that: In the spatiotemporal graph module, the attributes of the nodes include current flow, average speed and saturation. The spatiotemporal graph module reconstructs the graph structure at regular intervals to update the real-time status.

8. The traffic signal adaptive control system according to claim 1, characterized in that: The perception and transmission module uses the Kalman filter algorithm to perform drift correction on the high-precision positioning data collected by the vehicle terminal, and matches it with the high-precision map to lane-level accuracy.

9. A traffic signal adaptive control system according to claim 1 or 2, characterized in that: In the decision module, the state space includes the current phase combination, the number of vehicles in each direction, the saturation of each direction and the remaining time of the phase, and the action space includes the green light duration adjustment amount, the red light duration adjustment amount and the phase priority adjustment coefficient.

10. A traffic signal adaptive control method, based on a traffic signal adaptive control system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1, collects vehicle dynamic data, lane-level traffic parameters and historical traffic flow data in real time to build a perception network; S2, constructs a dynamic spatiotemporal graph based on the perception network and the current phase state; S3, generates signal control strategies based on the state data of the spatiotemporal graph; S4, execute the strategy and adjust the traffic light, and feed the execution effect back to the perception network for re-collection.

Images